端侧推理论文趋势 2023–2026

主要结论

• 端侧推理论文从 2023 年的 7 篇增长到 2026 年的 28 篇，三年间扩大到约 4.0 倍。
• NeurIPS 是发端侧推理论文最多的会议，2023–2026 年共 21 篇，占四个会议总量的 38%。
• 交叉方向里涨得最快的是联邦边缘推理：同时属于两个主题的论文从 2023 年的 3 篇增至 2026 年的 15 篇。
• 在「硬件与编译」的 5 个主题里，它的 2026 年论文数排第 4；同组增长最快的芯片设计 ML是 2.3 倍，端侧推理是 4.0 倍。

代表论文

按年份和会议均匀挑选，方便快速看到每一届的样貌

• NeurIPS 2026Does Sparse Connectivity Improve Generalization? Convolutional Networks Below the Edge of Stability
• ICML 2026Faster Activation Functions at the Edge for Post-Training Speedups
• ICLR 2026Neural Graduated Assignment for Maximum Common Edge Subgraphs
• CVPR 2026E2-SCI Elastic Edge-Cloud Speculative Decoding via Credit Inertia
• NeurIPS 2025CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment
• ICML 2025Edge-Colored Clustering in Hypergraphs: Beyond Minimizing Unsatisfied Edges
• ICLR 2025Grokking at the Edge of Numerical Stability
• CVPR 2025QuartDepth: Post-Training Quantization for Real-Time Depth Estimation on the Edge
• NeurIPS 2024Stepping on the Edge: Curvature Aware Learning Rate Tuners
• CVPR 20243D Neural Edge Reconstruction
• NeurIPS 2023From Trainable Negative Depth to Edge Heterophily in Graphs
• ICML 2023Second-order regression models exhibit progressive sharpening to the edge of stability
• ICLR 2023Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability
• CVPR 2023The Treasure Beneath Multiple Annotations: An Uncertainty-Aware Edge Detector
• NeurIPS 2026Edge of Stability Selectively Shapes Learning Across the Data Distribution
• ICML 2026PipeSD: An Efficient Cloud-Edge Collaborative Pipeline Inference Framework with Speculative Decoding
• ICLR 2026Tequila: Deadzone-free Ternary Quantization for Large Language Models
• CVPR 2026Edge-Focused Super-Resolution for Omnidirectional Images with Spherical Geometric Augmentation
• NeurIPS 2025Sim-LLM: Optimizing LLM Inference at the Edge through Inter-Task KV Reuse
• ICML 2025Towards characterizing the value of edge embeddings in Graph Neural Networks

全部端侧推理论文

共 55 篇，第 1 / 2 页

1. CVPR 2026E2-SCI Elastic Edge-Cloud Speculative Decoding via Credit Inertia
2. CVPR 2026Edge-Focused Super-Resolution for Omnidirectional Images with Spherical Geometric Augmentation
3. CVPR 2026ESAM Efficient Online 3D Perception on the Edge
4. CVPR 2026Foundry Distilling 3D Foundation Models for the Edge
5. CVPR 2026High-Fidelity Mobile Avatars with Pruned Local Blendshapes
6. CVPR 2026NanoSD Edge Efficient Foundation Model for Real Time Image Restoration
7. ICLR 2026Neural Graduated Assignment for Maximum Common Edge Subgraphs
8. ICLR 2026On the Convergence Direction of Gradient Descent
9. ICLR 2026Tequila: Deadzone-free Ternary Quantization for Large Language Models
10. ICML 2026Benchmarking at the Edge of Comprehension
11. ICML 2026Conflicting Biases at the Edge of Stability: Norm versus Sharpness Regularization
12. ICML 2026Edge-colored Clustering in Hypergraphs: A MaxECC Approximation
13. ICML 2026Faster Activation Functions at the Edge for Post-Training Speedups
14. ICML 2026Momentum Further Constrains Sharpness at the Edge of Stochastic Stability
15. ICML 2026Non-Euclidean Gradient Descent Operates at the Edge of Stability
16. ICML 2026PipeSD: An Efficient Cloud-Edge Collaborative Pipeline Inference Framework with Speculative Decoding
17. ICML 2026RADE: Random Add-Drop Edge as a Regularizer
18. ICML 2026XTransfer: Modality-Agnostic Few-Shot Model Transfer for Human Sensing at the Edge
19. NeurIPS 2026Computer Use at the Edge of the Statistical Precipice
20. NeurIPS 2026Does Sparse Connectivity Improve Generalization? Convolutional Networks Below the Edge of Stability
21. NeurIPS 2026EAT: Eviction-Aware Training for Long-Context LLM Inference on Edge Devices
22. NeurIPS 2026Edge of Stability Selectively Shapes Learning Across the Data Distribution
23. NeurIPS 2026Edge of Stochastic Stability: Revisiting the Edge of Stability for SGD
24. NeurIPS 2026Generalization at the Edge of Stability
25. NeurIPS 2026Geometry-Constrained Kolmogorov–Arnold Networks: Learning Edge Geometry via Banach Duality
26. NeurIPS 2026MedPsy: State‑of‑the‑Art Small Medical Language Models for Efficient Edge Deployment
27. NeurIPS 2026MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenario
28. NeurIPS 2026SGD at the Edge of Stability: The Stochastic Sharpness Gap
29. CVPR 2025EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features
30. CVPR 2025QuartDepth: Post-Training Quantization for Real-Time Depth Estimation on the Edge
31. ICLR 2025Grokking at the Edge of Numerical Stability
32. ICLR 2025Learning Dynamics of Deep Matrix Factorization Beyond the Edge of Stability
33. ICML 2025Edge-Colored Clustering in Hypergraphs: Beyond Minimizing Unsatisfied Edges
34. ICML 2025Grokking at the Edge of Linear Separability
35. ICML 2025Towards characterizing the value of edge embeddings in Graph Neural Networks
36. NeurIPS 2025CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction
37. NeurIPS 2025CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment
38. NeurIPS 2025On the Edge of Memorization in Diffusion Models
39. NeurIPS 2025Sim-LLM: Optimizing LLM Inference at the Edge through Inter-Task KV Reuse
40. NeurIPS 2025Structure-Aware Spectral Sparsification via Uniform Edge Sampling
41. NeurIPS 2025Transfer Learning on Edge Connecting Probability Estimation Under Graphon Model
42. NeurIPS 2025Variational Learning Finds Flatter Solutions at the Edge of Stability
43. CVPR 20243D Neural Edge Reconstruction
44. CVPR 2024MuGE: Multiple Granularity Edge Detection
45. CVPR 2024RepViT: Revisiting Mobile CNN From ViT Perspective
46. NeurIPS 2024Finding Transformer Circuits With Edge Pruning
47. NeurIPS 2024Stepping on the Edge: Curvature Aware Learning Rate Tuners
48. NeurIPS 2024The Edge-of-Reach Problem in Offline Model-Based Reinforcement Learning
49. CVPR 2023The Treasure Beneath Multiple Annotations: An Uncertainty-Aware Edge Detector
50. ICLR 2023A new characterization of the edge of stability based on a sharpness measure aware of batch gradient distribution

常见问题

2023–2026 年四大会议一共有多少篇端侧推理论文？

共 55 篇，按年份分别是 2023 年 7 篇、2024 年 6 篇、2025 年 14 篇、2026 年 28 篇。

NeurIPS 2026 有多少篇端侧推理论文？

NeurIPS 2026 有 10 篇端侧推理论文，上一届是 7 篇。

ICLR 2026 有多少篇端侧推理论文？

ICLR 2026 有 3 篇端侧推理论文，上一届是 2 篇。

ICML 2026 有多少篇端侧推理论文？

ICML 2026 有 9 篇端侧推理论文，上一届是 3 篇。

CVPR 2026 有多少篇端侧推理论文？

CVPR 2026 有 6 篇端侧推理论文，上一届是 2 篇。

哪个会议的端侧推理论文最多？

NeurIPS，2023–2026 年共 21 篇。